AI and Phenotypic screens lead to the discovery of common treatments for multiple distinct rare diseases with shared commonalities : A use case, and what's next

Importance: Over 300 million individuals worldwide are affected by rare diseases, with most lacking approved treatments. The exorbitant cost and complexities of traditional drug development necessitate orthogonal, expedited approaches for these conditions. Gap: Although AI has significantly impacted various sectors, including in non-rare drug discovery, their integration into rare disease drug discovery is still limited. Objective: Our research aims to advance the use of AI in rare diseases. To do so, our algorithms use weak signals from phenotypic drug screens on multiple distinct rare disease models in order to identify druggable targets that are shared by these diseases. These common targets serve as the basis for a second set of algorithms used to develop multifunctional drugs. Finally, the identified drugs are validated through the same phenotypic models used for target discovery, leading to a streamlined and cost-effective approach to drug development. Methods and Results: We validated our approach using subtypes of liposarcoma (LPS) as model diseases. LPS are a group of rare soft-tissue sarcomas, each with distinct pathological features, without treatment and with general poor prognosis. Prior screenings showed that models of the disease, irrespective of their type, display a high heterogeneity of response to being drugged. As such, limited drug candidates for further development in these diseases have been identified. Initial stages of our project involved using Kantify's AI target discovery algorithms to identify over 250 novel and shared targets in LPS, out of which 8 were shortlisted for further analysis. CRISPR-Cas9 knockdown of the shortlisted targets on two highly distinct LPS subtypes (myxoid, dedifferentiated) demonstrated the importance and dependence of all 8 of these targets in at least one of the two studied LPS models, and the commonality of dependence of 5 of the target in both lines. Subsequently, Kantify's AI drug discovery algorithms virtually screened 10,000 natural or repurposable compounds to identify over 80 possible treatments predicted to act on the identified shared targets. None of these 80 drugs had previously been screened in LPS. 8 (plus 1 negative control) drugs were selected for further validation in a viability assay on the two LPS models. 7 drugs were shown to have a highly similar and significant effect on both cell lines - 2 at low micromolar concentrations, and 5 at nanomolar concentrations. Implications: This research showcases the efficacy of AI to discover shared targets and multifunctional treatments for heterogeneous rare diseases, and highlights one promising path forward to make significant advancements for rare diseases. Currently, these algorithms are being further developed and validated in a variety of projects, including DREAMS (funded by Horizon Europe), an AI-based drug-discovery project focused on rare neuromuscular disorders, which leverages patient-derived induced pluripotent stem cells as phenotypic model.